Not the count. The split, and the time to kill.
The split
My Instant Prototype Agent turns a customer request into a working prototype on a preview URL in 7 to 10 minutes. Tracking where those prototypes went: about a third became the production feature after engineering cleaned them up, a third triggered a better idea than the one requested, and a third were rejected in 20 minutes.
The killed third is the one that paid for the tool. The old price of that no was a three-week spike.
Stripe's Harbor, as Rich Holmes wrote it up, has produced more than 12,000 prototypes since May. That headline treats all three outcomes as one unit. Report three columns instead: shipped, replaced by a better idea, killed.
The time to kill
Add one more number, the median time from first render to the kill decision. When prototypes are doing their job the kill comes fast, because finding out was the point. When that median climbs, prototypes are lingering, and a lingering prototype is no longer learning anything.
What a lingering prototype becomes
Production with no owner. Harbor's fastest growth is in finance, risk, and sales teams building internal dashboards they keep. Across the 39 agents I shipped in 80 days, 13 are orphaned, the best predictor of survival was a named human owner, and each survivor costs about two hours a month to keep true. A prototype that stops being thrown away needs a name and that budget, or it is the next thing that quietly stops being true.
The full argument is in Count the Kills, Not the Prototypes, part of AI Product Management.